Trends

Explore how topics appear across the podcast corpus. Each chunk of transcript is scored against your query using a sigmoid-weighted similarity function, producing a smooth relevance signal rather than a binary match.

By time

See how a topic's prevalence changes over time. Weeks are merged into adaptive bins so each data point rests on comparable evidence. The chart shows deviation from the median rate — above the line means more attention than usual.

By feed

See which podcast feeds are most (or least) relevant to a query. Every feed gets a relevance score based on the sum of sigmoid weights across all its chunks. Switch to deviation mode to see which feeds sit above or below the median.

How the scoring works

Each chunk's similarity to the query is passed through a logistic (sigmoid) function that maps it to a weight between 0 and 1. Two parameters control the shape: center (the similarity at which the weight is 0.5) and steepness (how sharply the function transitions). Both are tunable on the analysis pages. A high steepness approximates a hard threshold; a low steepness lets distant matches contribute fractional weight — useful for broad or vibes-based queries.